-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathutils.py
178 lines (144 loc) · 5.99 KB
/
utils.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
from sklearn.metrics import roc_auc_score, roc_curve
import torch
import numpy as np
import os
from sklearn.preprocessing import label_binarize
from sklearn.metrics import auc as calc_auc
def get_cam_1d(classifier, features):
tweight = list(classifier.parameters())[-2]
cam_maps = torch.einsum('bgf,cf->bcg', [features, tweight])
return cam_maps
def roc_threshold(label, prediction):
fpr, tpr, threshold = roc_curve(label, prediction, pos_label=1)
fpr_optimal, tpr_optimal, threshold_optimal = optimal_thresh(fpr, tpr, threshold)
c_auc = roc_auc_score(label, prediction)
return c_auc, threshold_optimal
def optimal_thresh(fpr, tpr, thresholds, p=0):
loss = (fpr - tpr) - p * tpr / (fpr + tpr + 1)
idx = np.argmin(loss, axis=0)
return fpr[idx], tpr[idx], thresholds[idx]
def eval_metric(oprob, label):
auc, threshold = roc_threshold(label.cpu().numpy(), oprob.detach().cpu().numpy())
prob = oprob > threshold
label = label ==1
TP = (prob & label).sum(0).float()
TN = ((~prob) & (~label)).sum(0).float()
FP = (prob & (~label)).sum(0).float()
FN = ((~prob) & label).sum(0).float()
# print("tp",TP)
# print("tn",TN)
# print("fp",FP)
# print("fn",FN)
# print("pred",prob)
# print("label",label)
accuracy = torch.mean(( TP + TN ) / ( TP + TN + FP + FN + 1e-12))
precision = torch.mean(TP / (TP + FP + 1e-12))
recall = torch.mean(TP / (TP + FN + 1e-12))
# print("recall",recall)
# print("precision",precision)
specificity = torch.mean( TN / (TN + FP + 1e-12))
F1 = 2*(precision * recall) / (precision + recall+1e-12)
return accuracy, precision, recall, specificity, F1, auc
def validate_clam(cur, epoch, model, loader, n_classes, early_stopping=None,loss_fn=None,
results_dir=None):
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.eval()
acc_logger = Accuracy_Logger(n_classes=n_classes)
inst_logger = Accuracy_Logger(n_classes=n_classes)
val_loss = 0.
val_error = 0.
val_inst_loss = 0.
val_inst_acc = 0.
inst_count = 0
prob = np.zeros((len(loader), n_classes))
labels = np.zeros(len(loader))
sample_size = model.k_sample
with torch.no_grad():
for batch_idx, (data, label) in enumerate(loader):
data, label = data.cuda(), label.cuda()
data=data.squeeze(0)
logits, Y_prob, Y_hat, _, instance_dict = model(data, label=label, instance_eval=True)
label=label.view([1])
acc_logger.log(Y_hat, label)
loss = loss_fn(logits, label)
val_loss += loss.item()
instance_loss = instance_dict['instance_loss']
inst_count += 1
instance_loss_value = instance_loss.item()
val_inst_loss += instance_loss_value
inst_preds = instance_dict['inst_preds']
inst_labels = instance_dict['inst_labels']
inst_logger.log_batch(inst_preds, inst_labels)
prob[batch_idx] = Y_prob.cpu().numpy()
labels[batch_idx] = label.item()
error = calculate_error(Y_hat, label)
val_error += error
val_error /= len(loader)
val_loss /= len(loader)
if n_classes == 2:
auc = roc_auc_score(labels, prob[:, 1])
aucs = []
else:
aucs = []
binary_labels = label_binarize(labels, classes=[i for i in range(n_classes)])
for class_idx in range(n_classes):
if class_idx in labels:
fpr, tpr, _ = roc_curve(binary_labels[:, class_idx], prob[:, class_idx])
aucs.append(calc_auc(fpr, tpr))
else:
aucs.append(float('nan'))
auc = np.nanmean(np.array(aucs))
print('\nVal Set, val_loss: {:.4f}, val_error: {:.4f}, auc: {:.4f}'.format(val_loss, val_error, auc))
if inst_count > 0:
val_inst_loss /= inst_count
for i in range(2):
acc, correct, count = inst_logger.get_summary(i)
print('class {} clustering acc {}: correct {}/{}'.format(i, acc, correct, count))
# if writer:
# writer.add_scalar('val/loss', val_loss, epoch)
# writer.add_scalar('val/auc', auc, epoch)
# writer.add_scalar('val/error', val_error, epoch)
# writer.add_scalar('val/inst_loss', val_inst_loss, epoch)
for i in range(n_classes):
acc, correct, count = acc_logger.get_summary(i)
print('class {}: acc {}, correct {}/{}'.format(i, acc, correct, count))
# if writer and acc is not None:
# writer.add_scalar('val/class_{}_acc'.format(i), acc, epoch)
if early_stopping:
assert results_dir
early_stopping(epoch, val_loss, model, ckpt_name=os.path.join(results_dir, "s_{}_checkpoint.pt".format(cur)))
if early_stopping.early_stop:
print("Early stopping")
return True
return False
def calculate_error(Y_hat, Y):
error = 1. - Y_hat.float().eq(Y.float()).float().mean().item()
return error
class Accuracy_Logger(object):
"""Accuracy logger"""
def __init__(self, n_classes):
super(Accuracy_Logger, self).__init__()
self.n_classes = n_classes
self.initialize()
def initialize(self):
self.data = [{"count": 0, "correct": 0} for i in range(self.n_classes)]
def log(self, Y_hat, Y):
Y_hat = int(Y_hat)
Y = int(Y)
self.data[Y]["count"] += 1
self.data[Y]["correct"] += (Y_hat == Y)
def log_batch(self, Y_hat, Y):
Y_hat = np.array(Y_hat).astype(int)
Y = np.array(Y).astype(int)
for label_class in np.unique(Y):
cls_mask = Y == label_class
self.data[label_class]["count"] += cls_mask.sum()
self.data[label_class]["correct"] += (Y_hat[cls_mask] == Y[cls_mask]).sum()
def get_summary(self, c):
count = self.data[c]["count"]
correct = self.data[c]["correct"]
if count == 0:
acc = None
else:
acc = float(correct) / count
return acc, correct, count